Adding auxiliary left and right phoneme prediction heads to full-sum ASR training improves word error rates, especially on 300h Switchboard, and enables full-sum-only factored hybrid HMM training without external alignments.
Monotonic recu rrent neural network transducer and decoding strategies,
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.SD 1years
2025 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Right Label Context in End-to-End Training of Time-Synchronous ASR Models
Adding auxiliary left and right phoneme prediction heads to full-sum ASR training improves word error rates, especially on 300h Switchboard, and enables full-sum-only factored hybrid HMM training without external alignments.